PragyaDoc: A Universal Document Intelligence Framework for Multilingual Medical Document Understanding in Low-Resource Settings
Author: Jagpal Singh Jhala
Submitted: 1 June 2026 (v1)
Subjects: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
arXiv ID: 2608.07478
DOI: https://doi.org/10.48550/arXiv.2608.07478
Abstract
India's 22 official languages present a significant accessibility challenge in healthcare: the vast majority of medical documentation is produced exclusively in English, yet the individuals who most urgently need this information—rural populations, ASHA (Accredited Social Health Activist) workers, and patients' families—are effectively excluded from understanding it. This paper introduces PragyaDoc, a Universal Document Intelligence Framework designed to bridge this gap. The framework employs a four-layer pipeline: (1) a parallel ensemble OCR extraction layer, (2) a geometric-lexical fusion layer, (3) a deterministic domain structuring layer, and (4) a dual-LLM medical reasoning and localization layer. Together, these components enable accurate, context-aware translation and interpretation of medical documents across India's linguistic landscape, even in settings with limited computational resources.
1. Introduction
In 2026, the Indian healthcare system faces a paradox: while medical knowledge grows at an unprecedented rate, the ability to disseminate that knowledge to those who need it remains constrained. English remains the dominant language in medical records, pharmaceutical instructions, and clinical guidelines. This poses a critical barrier for millions of Hindi, Tamil, Bengali, and other regional language speakers who interact with healthcare systems daily.
Current solutions rely on generic machine translation tools, which have demonstrated significant limitations when applied to technical medical content. These tools often fail to capture domain-specific terminology, struggle with the structural complexity of medical documents, and lack the contextual reasoning required to ensure patient safety.
PragyaDoc addresses these limitations through a purpose-built architecture that combines advances in optical character recognition (OCR), geometric document analysis, and large language model (LLM) reasoning—all designed to function effectively in low-resource settings commonly found across rural India.
2. The Accessibility Problem
India's linguistic diversity, while culturally rich, creates substantial operational challenges in healthcare delivery. Key observations:
- Document asymmetry: Over 90% of medical documentation is generated in English, while only roughly 10% of the population speaks English fluently.
- Critical use cases: Prescription instructions, discharge summaries, and diagnostic reports all require comprehension by patients and their families.
- Frontline worker needs: ASHA workers—often the first point of medical contact for rural populations—require accessible medical information in local languages.
- Safety implications: Misunderstanding medical instructions can lead to medication errors, adverse reactions, and poor treatment adherence.
3. Framework Architecture
The PragyaDoc framework consists of four integrated layers that work in concert to deliver accurate, multilingual medical document understanding.
3.1 Parallel Ensemble OCR Extraction Layer
This layer employs multiple OCR engines operating in parallel, each optimized for different document types and quality levels. By combining outputs through a voting mechanism, the framework achieves robust text extraction even from degraded or low-quality scans commonly encountered in field settings.
Key features:
- Support for multiple Indian scripts and mixed-script documents
- Adaptive preprocessing for varying image qualities
- Confidence scoring for each extracted element
3.2 Geometric-Lexical Fusion Layer
The second layer performs spatial analysis of document layout, mapping the relationships between headings, body text, tables, and annotations. This geometric understanding is then fused with lexical information to preserve the document's structural integrity during translation.
This layer addresses a critical gap in conventional translation systems: the ability to maintain context based on document position. For example, distinguishing between a drug name in a heading versus a body-text mention requires understanding both layout and semantics.
3.3 Deterministic Domain Structuring Layer
The third layer applies medical domain ontologies and structural rules to normalize the extracted content. This deterministic approach ensures medical terminology consistency, validates that critical information fields are present, and structures content for downstream processing.
Key components:
- Named entity recognition for medications, dosages, and diagnoses
- Validation rules for dosage ranges and drug interactions
- Standard medical vocabulary mapping (ICD, SNOMED CT)
3.4 Dual-LLM Medical Reasoning and Localization Layer
The final layer employs two complementary large language models:
- The Reasoning LLM interprets the medical content structurally, ensuring domain accuracy and clinical safety.
- The Localization LLM performs culturally appropriate translation, adapting content for regional dialects and health literacy levels.
- Model quantization for edge device deployment
- Offline operation capabilities for regions with limited connectivity
- Progressive enhancement when better hardware becomes available
- Patient discharge summaries: Translating key instructions into patient-understandable language
- Prescription instructions: Converting complex dosing information into clear local language guidance
- Community health worker tools: Enabling ASHA workers to retrieve and understand medical guidelines in relevant languages
- Informed consent documents: Ensuring patients genuinely understand procedures before signing
- Full-scale experimental validation across Indian medical datasets
- Comparative evaluation against existing translation systems
- Deployment trials in rural healthcare settings
- Expansion beyond Indian languages to other multilingual regions
This division of labor allows the framework to balance technical precision with accessibility—a critical consideration when translating between medical English and colloquial regional languages.
For deployment in low-resource settings, the framework supports:
4. Significance and Use Cases
PragyaDoc has direct applications across multiple healthcare scenarios:
In addition to translation, the framework's localization dimension can generate visual materials and simplified text for patients with low literacy. It also includes accessibility features for individuals with visual impairments.
5. Current Status and Outlook
As of the June 2026 submission, PragyaDoc is presented as a research framework with a complete architectural specification. Future work will address:
6. Conclusion
PragyaDoc represents a significant step toward equitable healthcare access in linguistically diverse regions. By combining state-of-the-art document intelligence techniques with a pragmatic focus on low-resource settings, the framework offers a viable path to bridging the medical language gap in India and beyond.
The framework's four-layer architecture—OCR extraction, geometric-lexical fusion, domain structuring, and dual-LLM reasoning—provides a comprehensive solution that prioritizes both clinical accuracy and patient comprehension. As the system progresses from architectural proposal to validated implementation, it holds the potential to transform how medical knowledge is disseminated across India's multilingual landscape.
This paper is available under a Creative Commons license. Full text access: HTML (experimental) via arXiv.
via ArXiv CV